Claude Code's boss: a TUI that enforces gates, budgets, and worktrees

What it does

Foreman is a keyboard-driven terminal UI that spawns the local claude CLI in headless mode and marches it through a rigid software-delivery pipeline: plan, design docs, issue slicing, TDD build, and end-to-end tests. Human reviewers must approve each design phase before the agent is allowed to proceed to the next, and every bit of state lives as human-readable files committed inside the target repository.

The interesting bit

The guardrails are enforced by the orchestrator, not the agent. Foreman caps per-run turns, costs, and time; imposes a daily cost ceiling with a hard stop; isolates parallel workers in separate git worktrees; and uses a PreToolUse hook to stop agents from writing their own verification. If the process dies mid-build, it recovers entirely from disk because there is no database.

Key highlights

• Gated pipeline (plan → ADR/PRD → issues → TDD build → e2e) with hash-sealed approvals that auto-invalidate if a document is edited after sign-off.
• Parses the claude CLI’s stream-json event stream directly and enforces budgets externally.
• All state is repo-committed files; crash the TUI and resume later with zero data loss.
• An evals flywheel (foreman retro and foreman bench) clusters recurring failures and gates prompt patches until they improve success rates.
• Keyboard-driven Textual TUI with a demo mode that runs the full pipeline against a mocked backend without spending tokens.

Caveats

• Requires the locally installed claude CLI and only runs on Linux or WSL2; it is not a standalone agent.
• The README notes the project is developed and tested on Ubuntu under WSL2, so mileage may vary on other platforms.

Verdict

Worth a look if you are already using Claude Code and want external guardrails, cost controls, and a human-in-the-loop delivery flow. Skip it if you need a fully autonomous agent or are not prepared to babysit the design phases.